create_history_aware_retriever with a single retriever node

[Abraham-Martin-at-ANYVERSE]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

The create_history_aware_retriever function, which can be imported like this:

from langchain.chains.history_aware_retriever import create_history_aware_retriever

has a problem that makes TruLens evaluation tests for RAG context documents unusable for LangChain pipelines: it adds a retriever two times to the chain, so TruLens cannot find out which one to use for evaluations related to context documents. Some example code:

    history_aware_retriever = create_history_aware_retriever(
        llm,
        base_retriever,
        contextualize_q_prompt
    ) 
    question_answer_chain = create_stuff_documents_chain(llm, chat_prompt_template)
    rag_chain= create_retrieval_chain(
        history_aware_retriever,
        question_answer_chain
    )
    context = TruChain.select_context(rag_chain)

    feedbacks.append(
        Feedback(provider.context_relevance_with_cot_reasons,
                name="Context Relevance")
        .on_input()
        .on(context)
        .aggregate(np.mean)
    )

Running this through TruLens gives the following error:

ValueError: Found more than one `BaseRetriever` in app:
        <class 'langchain_core.vectorstores.base.VectorStoreRetriever'> at first.bound.first.mapper.steps__.context.bound.branches[0][1].last
        <class 'langchain_core.vectorstores.base.VectorStoreRetriever'> at first.bound.first.mapper.steps__.context.bound.default.last

I understand that TruLens can make things better, but I think that LangChain can make things easier not for TruLens, but for other tools that can have the same issue assuming pipelines with only one retriever.
So, I ask for chaging create_history_aware_retriever in a way that TruLens, and eventually other tools, doesn't fail.

Motivation

TruLens doesn't work for context-retrieved documents' evaluations when create_history_aware_retriever is used

Proposal (If applicable)

I propose to change the implementation of create_history_aware_retriever from this:

    if "input" not in prompt.input_variables:
        raise ValueError(
            "Expected `input` to be a prompt variable, "
            f"but got {prompt.input_variables}"
        )

    retrieve_documents: RetrieverOutputLike = RunnableBranch(
        (
            # Both empty string and empty list evaluate to False
            lambda x: not x.get("chat_history", False),
            # If no chat history, then we just pass input to retriever
            (lambda x: x["input"]) | retriever,
        ),
        # If chat history, then we pass inputs to LLM chain, then to retriever
        prompt | llm | StrOutputParser() | retriever,
    ).with_config(run_name="chat_retriever_chain")
    return retrieve_documents

to this:

    if "input" not in prompt.input_variables:
        raise ValueError(
            "Expected `input` to be a prompt variable, "
            f"but got {prompt.input_variables}"
        )
    pre_chain: RetrieverOutputLike = RunnableBranch(
        (
            # Both empty string and empty list evaluate to False
            lambda x: not x.get("chat_history", False),
            # If no chat history, then we just pass input to retriever
            (lambda x: x["input"]),
        ),
        # If chat history, then we pass inputs to LLM chain, then to retriever
        prompt | llm | StrOutputParser(),
    )
    retrieve_documents: RetrieverOutputLike = (
        pre_chain | retriever
    ).with_config(run_name="chat_retriever_chain")
    return retrieve_documents

Where we only have one retriever node, which is less error-prone for external tools such as TruLens.